Real-time hazard detection powered by AI and Snowflake.
ThreeOneOne is a real-time hazard detection platform that aggregates posts from social media (X, Bluesky, Reddit) and direct user submissions, turning raw, noisy data into verified, actionable incidents on a live map.
Devpost: https://devpost.com/software/threeoneone-nrfyic
I casually brought up calling 311 and was astonished when none of my team members had heard of it.
- The Problem: The 311 Toronto app is used by less than 20% of residents, and only 5% of university students even know it exists.
- The Gap: While 311 calls lag, hazards like fallen trees or broken signals appear on social media within minutes. However, cities can't act on unverified tweets.
- The Solution: ThreeOneOne acts as the "intelligence layer" in between—listening to public signals, using AI to decide what's a hazard, and clustering reports to surface incidents only when there is enough corroborating evidence to act.
ThreeOneOne turns raw text and images into verified incidents on an interactive map.
- Centralize: All posts flow into Snowflake, our single source of truth.
- Classify: Uses Snowflake Cortex and Gemini 1.5 Flash-Lite to filter noise and categorize hazards.
- Vision & Reasoning: Gemini 1.5 Flash performs multimodal analysis to confirm hazards in images, while Grounded Search pins vague locations (like "near the bridge") to precise GPS coordinates.
- Cluster & Score: DBSCAN groups reports by location and time. Each incident gets a confidence score based on report count, source diversity, and image evidence.
- Verify: High-confidence clusters are cross-referenced against local news sources to earn a "Confirmed" badge.
| Layer | Technologies |
|---|---|
| Frontend | React, TypeScript, Vite, Mapbox GL JS |
| Backend | FastAPI, Python |
| Data & AI | Snowflake (Geospatial Lake), Snowflake Cortex |
| Models | Gemini 3.1 Flash Lite, Gemini 2.5 Flash (Grounded Search & Vision) |
| Analytics | DBSCAN, Haversine Metric |
Snowflake serves as our Geospatial Data Lake. We ingest raw JSON via VARIANT columns and use native Geospatial functions (like ST_GEOHASH) to turn noise into mapped data. Cortex AI summarizes long posts into one-sentence "Dispatch Notes" for city workers.
We utilize a Tiered Inference Pipeline for cost-effective scaling:
- Gatekeeper: Flash-Lite filters out irrelevant posts.
- Multimodal: Flash performs spatial reasoning to identify landmarks and signs to infer GPS coordinates where metadata is missing.
- Validation: Enriched data is pushed back to Snowflake to power the real-time map.
- Explainable AI: We built a custom confidence formula (
scoring.py) that makes the AI's decision-making transparent to users and responders. - Production-Ready: By putting Snowflake at the center from Day 1, our data model is built for scale, not just a static demo.
- Grounded Metadata: We learned to leverage
grounding_chunksto recover news URLs even when raw JSON responses were malformed. - Spatial-Temporal Clustering: Discovered that a two-pass approach (Spatial first, then Temporal) creates much cleaner incident clusters than a single unified metric.
- More Social Ingest: Integrating Meta and Instagram APIs for broader data coverage.
- Admin Dashboard: Building Snowflake-side analytics to track hazard density by neighborhood.
- Live Updates: Implementing WebSockets to push new clusters to the map instantly.
- City Partnerships: Piloting the platform as a "pre-call" signal feed for municipal 311 teams.